Mike Preuss

Paula Torren-Peraire, Alan Kai Hassen, Samuel Genheden et al.

Retrosynthesis consists of breaking down a chemical compound recursively step-by-step into molecular precursors until a set of commercially available molecules is found with the goal to provide a synthesis route. Its two primary research directions, single-step retrosynthesis prediction, which models the chemical reaction logic, and multi-step synthesis planning, which tries to find the correct sequence of reactions, are inherently intertwined. Still, this connection is not reflected in contemporary research. In this work, we combine these two major research directions by applying multiple single-step retrosynthesis models within multi-step synthesis planning and analyzing their impact using public and proprietary reaction data. We find a disconnection between high single-step performance and potential route-finding success, suggesting that single-step models must be evaluated within synthesis planning in the future. Furthermore, we show that the commonly used single-step retrosynthesis benchmark dataset USPTO-50k is insufficient as this evaluation task does not represent model performance and scalability on larger and more diverse datasets. For multi-step synthesis planning, we show that the choice of the single-step model can improve the overall success rate of synthesis planning by up to +28% compared to the commonly used baseline model. Finally, we show that each single-step model finds unique synthesis routes, and differs in aspects such as route-finding success, the number of found synthesis routes, and chemical validity, making the combination of single-step retrosynthesis prediction and multi-step synthesis planning a crucial aspect when developing future methods.

Alan Kai Hassen, Paula Torren-Peraire, Samuel Genheden et al.

Retrosynthesis is the task of breaking down a chemical compound recursively step-by-step into molecular precursors until a set of commercially available molecules is found. Consequently, the goal is to provide a valid synthesis route for a molecule. As more single-step models develop, we see increasing accuracy in the prediction of molecular disconnections, potentially improving the creation of synthetic paths. Multi-step approaches repeatedly apply the chemical information stored in single-step retrosynthesis models. However, this connection is not reflected in contemporary research, fixing either the single-step model or the multi-step algorithm in the process. In this work, we establish a bridge between both tasks by benchmarking the performance and transfer of different single-step retrosynthesis models to the multi-step domain by leveraging two common search algorithms, Monte Carlo Tree Search and Retro*. We show that models designed for single-step retrosynthesis, when extended to multi-step, can have a tremendous impact on the route finding capabilities of current multi-step methods, improving performance by up to +30% compared to the most widely used model. Furthermore, we observe no clear link between contemporary single-step and multi-step evaluation metrics, showing that single-step models need to be developed and tested for the multi-step domain and not as an isolated task to find synthesis routes for molecules of interest.

Marco Pleines, Matthias Pallasch, Frank Zimmer et al.

At first sight it may seem straightforward to use recurrent layers in Deep Reinforcement Learning algorithms to enable agents to make use of memory in the setting of partially observable environments. Starting from widely used Proximal Policy Optimization (PPO), we highlight vital details that one must get right when adding recurrence to achieve a correct and efficient implementation, namely: properly shaping the neural net's forward pass, arranging the training data, correspondingly selecting hidden states for sequence beginnings and masking paddings for loss computation. We further explore the limitations of recurrent PPO by benchmarking the contributed novel environments Mortar Mayhem and Searing Spotlights that challenge the agent's memory beyond solely capacity and distraction tasks. Remarkably, we can demonstrate a transition to strong generalization in Mortar Mayhem when scaling the number of training seeds, while the agent does not succeed on Searing Spotlights, which seems to be a tough challenge for memory-based agents.

Marco Pleines, Matthias Pallasch, Frank Zimmer et al.

Memory Gym presents a suite of 2D partially observable environments, namely Mortar Mayhem, Mystery Path, and Searing Spotlights, designed to benchmark memory capabilities in decision-making agents. These environments, originally with finite tasks, are expanded into innovative, endless formats, mirroring the escalating challenges of cumulative memory games such as "I packed my bag". This progression in task design shifts the focus from merely assessing sample efficiency to also probing the levels of memory effectiveness in dynamic, prolonged scenarios. To address the gap in available memory-based Deep Reinforcement Learning baselines, we introduce an implementation within the open-source CleanRL library that integrates Transformer-XL (TrXL) with Proximal Policy Optimization. This approach utilizes TrXL as a form of episodic memory, employing a sliding window technique. Our comparative study between the Gated Recurrent Unit (GRU) and TrXL reveals varied performances across our finite and endless tasks. TrXL, on the finite environments, demonstrates superior effectiveness over GRU, but only when utilizing an auxiliary loss to reconstruct observations. Notably, GRU makes a remarkable resurgence in all endless tasks, consistently outperforming TrXL by significant margins. Website and Source Code: https://marcometer.github.io/jmlr_2024.github.io/

Marco Pleines, Konstantin Ramthun, Yannik Wegener et al.

Autonomously trained agents that are supposed to play video games reasonably well rely either on fast simulation speeds or heavy parallelization across thousands of machines running concurrently. This work explores a third way that is established in robotics, namely sim-to-real transfer, or if the game is considered a simulation itself, sim-to-sim transfer. In the case of Rocket League, we demonstrate that single behaviors of goalies and strikers can be successfully learned using Deep Reinforcement Learning in the simulation environment and transferred back to the original game. Although the implemented training simulation is to some extent inaccurate, the goalkeeping agent saves nearly 100% of its faced shots once transferred, while the striking agent scores in about 75% of cases. Therefore, the trained agent is robust enough and able to generalize to the target domain of Rocket League.

Zhao Yang, Thomas. M. Moerland, Mike Preuss et al.

While deep reinforcement learning has shown important empirical success, it tends to learn relatively slow due to slow propagation of rewards information and slow update of parametric neural networks. Non-parametric episodic memory, on the other hand, provides a faster learning alternative that does not require representation learning and uses maximum episodic return as state-action values for action selection. Episodic memory and reinforcement learning both have their own strengths and weaknesses. Notably, humans can leverage multiple memory systems concurrently during learning and benefit from all of them. In this work, we propose a method called Two-Memory reinforcement learning agent (2M) that combines episodic memory and reinforcement learning to distill both of their strengths. The 2M agent exploits the speed of the episodic memory part and the optimality and the generalization capacity of the reinforcement learning part to complement each other. Our experiments demonstrate that the 2M agent is more data efficient and outperforms both pure episodic memory and pure reinforcement learning, as well as a state-of-the-art memory-augmented RL agent. Moreover, the proposed approach provides a general framework that can be used to combine any episodic memory agent with other off-policy reinforcement learning algorithms.

Zhao Yang, Thomas M. Moerland, Mike Preuss et al.

Go-Explore achieved breakthrough performance on challenging reinforcement learning (RL) tasks with sparse rewards. The key insight of Go-Explore was that successful exploration requires an agent to first return to an interesting state ('Go'), and only then explore into unknown terrain ('Explore'). We refer to such exploration after a goal is reached as 'post-exploration'. In this paper, we present a clear ablation study of post-exploration in a general intrinsically motivated goal exploration process (IMGEP) framework, that the Go-Explore paper did not show. We study the isolated potential of post-exploration, by turning it on and off within the same algorithm under both tabular and deep RL settings on both discrete navigation and continuous control tasks. Experiments on a range of MiniGrid and Mujoco environments show that post-exploration indeed helps IMGEP agents reach more diverse states and boosts their performance. In short, our work suggests that RL researchers should consider to use post-exploration in IMGEP when possible since it is effective, method-agnostic and easy to implement.

Matthias Müller-Brockhausen, Aske Plaat, Mike Preuss

Reinforcement Learning (RL) is one of the most dynamic research areas in Game AI and AI as a whole, and a wide variety of games are used as its prominent test problems. However, it is subject to the replicability crisis that currently affects most algorithmic AI research. Benchmarking in Reinforcement Learning could be improved through verifiable results. There are numerous benchmark environments whose scores are used to compare different algorithms, such as Atari. Nevertheless, reviewers must trust that figures represent truthful values, as it is difficult to reproduce an exact training curve. We propose improving this situation by providing access to the original experimental data to validate study results. To that end, we rely on the concept of minimal traces. These allow re-simulation of action sequences in deterministic RL environments and, in turn, enable reviewers to verify, re-use, and manually inspect experimental results without needing large compute clusters. It also permits validation of presented reward graphs, an inspection of individual episodes, and re-use of result data (baselines) for proper comparison in follow-up papers. We offer plug-and-play code that works with Gym so that our measures fit well in the existing RL and reproducibility eco-system. Our approach is freely available, easy to use, and adds minimal overhead, as minimal traces allow a data compression ratio of up to $\approx 10^4:1$ (94GB to 8MB for Atari Pong) compared to a regular MDP trace used in offline RL datasets. The paper presents proof-of-concept results for a variety of games.

Zhao Yang, Thomas M. Moerland, Mike Preuss et al.

Reinforcement learning (RL) is an appealing paradigm for training intelligent agents, enabling policy acquisition from the agent's own autonomously acquired experience. However, the training process of RL is far from automatic, requiring extensive human effort to reset the agent and environments. To tackle the challenging reset-free setting, we first demonstrate the superiority of model-based (MB) RL methods in such setting, showing that a straightforward adaptation of MBRL can outperform all the prior state-of-the-art methods while requiring less supervision. We then identify limitations inherent to this direct extension and propose a solution called model-based reset-free (MoReFree) agent, which further enhances the performance. MoReFree adapts two key mechanisms, exploration and policy learning, to handle reset-free tasks by prioritizing task-relevant states. It exhibits superior data-efficiency across various reset-free tasks without access to environmental reward or demonstrations while significantly outperforming privileged baselines that require supervision. Our findings suggest model-based methods hold significant promise for reducing human effort in RL. Website: https://yangzhao-666.github.io/morefree

Arthur van der Staaij, Jelmer Prins, Vincent L. Prins et al.

Procedural city generation that focuses on believability and adaptability to random terrain is a difficult challenge in the field of Procedural Content Generation (PCG). Dozens of researchers compete for a realistic approach in challenges such as the Generative Settlement Design in Minecraft (GDMC), in which our method has won the 2022 competition. This was achieved through a decentralised, iterative planning process that is transferable to similar generation processes that aims to produce "organic" content procedurally.

Maria Laura Santoni, Elena Raponi, Aneta Neumann et al.

In real-world applications, users often favor structurally diverse design choices over one high-quality solution. It is hence important to consider more solutions that decision makers can compare and further explore based on additional criteria. Alongside the existing approaches of evolutionary diversity optimization, quality diversity, and multimodal optimization, this paper presents a fresh perspective on this challenge by considering the problem of identifying a fixed number of solutions with a pairwise distance above a specified threshold while maximizing their average quality. We obtain first insight into these objectives by performing a subset selection on the search trajectories of different well-established search heuristics, whether they have been specifically designed with diversity in mind or not. We emphasize that the main goal of our work is not to present a new algorithm but to understand the capability of off-the-shelf algorithms to quantify the trade-off between the minimum pairwise distance within batches of solutions and their average quality. We also analyze how this trade-off depends on the properties of the underlying optimization problem. A possibly surprising outcome of our empirical study is the observation that naive uniform random sampling establishes a very strong baseline for our problem, hardly ever outperformed by the search trajectories of the considered heuristics. We interpret these results as a motivation to develop algorithms tailored to produce diverse solutions of high average quality.

Marco Pleines, Daniel Addis, David Rubinstein et al.

Pokémon Red, a classic Game Boy JRPG, presents significant challenges as a testbed for agents, including multi-tasking, long horizons of tens of thousands of steps, hard exploration, and a vast array of potential policies. We introduce a simplistic environment and a Deep Reinforcement Learning (DRL) training methodology, demonstrating a baseline agent that completes an initial segment of the game up to completing Cerulean City. Our experiments include various ablations that reveal vulnerabilities in reward shaping, where agents exploit specific reward signals. We also discuss limitations and argue that games like Pokémon hold strong potential for future research on Large Language Model agents, hierarchical training algorithms, and advanced exploration methods. Source Code: https://github.com/MarcoMeter/neroRL/tree/poke_red

Hui Wang, Michael Emmerich, Mike Preuss et al.

Since AlphaGo and AlphaGo Zero have achieved breakground successes in the game of Go, the programs have been generalized to solve other tasks. Subsequently, AlphaZero was developed to play Go, Chess and Shogi. In the literature, the algorithms are explained well. However, AlphaZero contains many parameters, and for neither AlphaGo, AlphaGo Zero nor AlphaZero, there is sufficient discussion about how to set parameter values in these algorithms. Therefore, in this paper, we choose 12 parameters in AlphaZero and evaluate how these parameters contribute to training. We focus on three objectives~(training loss, time cost and playing strength). For each parameter, we train 3 models using 3 different values~(minimum value, default value, maximum value). We use the game of play 6$\times$6 Othello, on the AlphaZeroGeneral open source re-implementation of AlphaZero. Overall, experimental results show that different values can lead to different training results, proving the importance of such a parameter sweep. We categorize these 12 parameters into time-sensitive parameters and time-friendly parameters. Moreover, through multi-objective analysis, this paper provides an insightful basis for further hyper-parameter optimization.

Qianpu Chen, Giulio Barbero, Mike Preuss et al.

We introduce Emergent Trust Learning (ETL), a lightweight, trust-based control algorithm that can be plugged into existing AI agents. It enables these to reach cooperation in competitive game environments under shared resources. Each agent maintains a compact internal trust state, which modulates memory, exploration, and action selection. ETL requires only individual rewards and local observations and incurs negligible computational and communication overhead. We evaluate ETL in three environments: In a grid-based resource world, trust-based agents reduce conflicts and prevent long-term resource depletion while achieving competitive individual returns. In a hierarchical Tower environment with strong social dilemmas and randomised floor assignments, ETL sustains high survival rates and recovers cooperation even after extended phases of enforced greed. In the Iterated Prisoner's Dilemma, the algorithm generalises to a strategic meta-game, maintaining cooperation with reciprocal opponents while avoiding long-term exploitation by defectors. Code will be released upon publication.

Aske Plaat, Max van Duijn, Niki van Stein et al.

There is great interest in agentic LLMs, large language models that act as agents. We review the growing body of work in this area and provide a research agenda. Agentic LLMs are LLMs that (1) reason, (2) act, and (3) interact. We organize the literature according to these three categories. The research in the first category focuses on reasoning, reflection, and retrieval, aiming to improve decision making; the second category focuses on action models, robots, and tools, aiming for agents that act as useful assistants; the third category focuses on multi-agent systems, aiming for collaborative task solving and simulating interaction to study emergent social behavior. We find that works mutually benefit from results in other categories: retrieval enables tool use, reflection improves multi-agent collaboration, and reasoning benefits all categories. We discuss applications of agentic LLMs and provide an agenda for further research. Important applications are in medical diagnosis, logistics and financial market analysis. Meanwhile, self-reflective agents playing roles and interacting with one another augment the process of scientific research itself. Further, agentic LLMs may provide a solution for the problem of LLMs running out of training data: inference-time behavior generates new training states, such that LLMs can keep learning without needing ever larger datasets. We note that there is risk associated with LLM assistants taking action in the real world, while agentic LLMs are also likely to benefit society.

Zhao Yang, Thomas M. Moerland, Mike Preuss et al.

Non-parametric episodic memory can be used to quickly latch onto high-rewarded experience in reinforcement learning tasks. In contrast to parametric deep reinforcement learning approaches in which reward signals need to be back-propagated slowly, these methods only need to discover the solution once, and may then repeatedly solve the task. However, episodic control solutions are stored in discrete tables, and this approach has so far only been applied to discrete action space problems. Therefore, this paper introduces Continuous Episodic Control (CEC), a novel non-parametric episodic memory algorithm for sequential decision making in problems with a continuous action space. Results on several sparse-reward continuous control environments show that our proposed method learns faster than state-of-the-art model-free RL and memory-augmented RL algorithms, while maintaining good long-run performance as well. In short, CEC can be a fast approach for learning in continuous control tasks.

Zhao Yang, Thomas M. Moerland, Mike Preuss et al.

Learning diverse skills without hand-crafted reward functions could accelerate reinforcement learning in downstream tasks. However, existing skill discovery methods focus solely on maximizing the diversity of skills without considering human preferences, which leads to undesirable behaviors and possibly dangerous skills. For instance, a cheetah robot trained using previous methods learns to roll in all directions to maximize skill diversity, whereas we would prefer it to run without flipping or entering hazardous areas. In this work, we propose a Foundation model Guided (FoG) skill discovery method, which incorporates human intentions into skill discovery through foundation models. Specifically, FoG extracts a score function from foundation models to evaluate states based on human intentions, assigning higher values to desirable states and lower to undesirable ones. These scores are then used to re-weight the rewards of skill discovery algorithms. By optimizing the re-weighted skill discovery rewards, FoG successfully learns to eliminate undesirable behaviors, such as flipping or rolling, and to avoid hazardous areas in both state-based and pixel-based tasks. Interestingly, we show that FoG can discover skills involving behaviors that are difficult to define. Interactive visualisations are available from https://sites.google.com/view/submission-fog.

Alan Kai Hassen, Andrius Bernatavicius, Antonius P. A. Janssen et al.

Applications of machine learning in chemistry are often limited by the scarcity and expense of labeled data, restricting traditional supervised methods. In this work, we introduce a framework for molecular reasoning using general-purpose Large Language Models (LLMs) that operates without requiring labeled training data. Our method anchors chain-of-thought reasoning to the molecular structure by using unique atomic identifiers. First, the LLM performs a one-shot task to identify relevant fragments and their associated chemical labels or transformation classes. In an optional second step, this position-aware information is used in a few-shot task with provided class examples to predict the chemical transformation. We apply our framework to single-step retrosynthesis, a task where LLMs have previously underperformed. Across academic benchmarks and expert-validated drug discovery molecules, our work enables LLMs to achieve high success rates in identifying chemically plausible reaction sites ($\geq90\%$), named reaction classes ($\geq40\%$), and final reactants ($\geq74\%$). Beyond solving complex chemical tasks, our work also provides a method to generate theoretically grounded synthetic datasets by mapping chemical knowledge onto the molecular structure and thereby addressing data scarcity.

Zhao Yang, Thomas M. Moerland, Mike Preuss et al.

Go-Explore achieved breakthrough performance on challenging reinforcement learning (RL) tasks with sparse rewards. The key insight of Go-Explore was that successful exploration requires an agent to first return to an interesting state ('Go'), and only then explore into unknown terrain ('Explore'). We refer to such exploration after a goal is reached as 'post-exploration'. In this paper we present a systematic study of post-exploration, answering open questions that the Go-Explore paper did not answer yet. First, we study the isolated potential of post-exploration, by turning it on and off within the same algorithm. Subsequently, we introduce new methodology to adaptively decide when to post-explore and for how long to post-explore. Experiments on a range of MiniGrid environments show that post-exploration indeed boosts performance (with a bigger impact than tuning regular exploration parameters), and this effect is further enhanced by adaptively deciding when and for how long to post-explore. In short, our work identifies adaptive post-exploration as a promising direction for RL exploration research.

Zhao Yang, Mike Preuss, Aske Plaat

Learning to solve sparse-reward reinforcement learning problems is difficult, due to the lack of guidance towards the goal. But in some problems, prior knowledge can be used to augment the learning process. Reward shaping is a way to incorporate prior knowledge into the original reward function in order to speed up the learning. While previous work has investigated the use of expert knowledge to generate potential functions, in this work, we study whether we can use a search algorithm(A*) to automatically generate a potential function for reward shaping in Sokoban, a well-known planning task. The results showed that learning with shaped reward function is faster than learning from scratch. Our results indicate that distance functions could be a suitable function for Sokoban. This work demonstrates the possibility of solving multiple instances with the help of reward shaping. The result can be compressed into a single policy, which can be seen as the first phrase towards training a general policy that is able to solve unseen instances.

Aske Plaat, Walter Kosters, Mike Preuss

Deep reinforcement learning has shown remarkable success in the past few years. Highly complex sequential decision making problems from game playing and robotics have been solved with deep model-free methods. Unfortunately, the sample complexity of model-free methods is often high. To reduce the number of environment samples, model-based reinforcement learning creates an explicit model of the environment dynamics. Achieving high model accuracy is a challenge in high-dimensional problems. In recent years, a diverse landscape of model-based methods has been introduced to improve model accuracy, using methods such as uncertainty modeling, model-predictive control, latent models, and end-to-end learning and planning. Some of these methods succeed in achieving high accuracy at low sample complexity, most do so either in a robotics or in a games context. In this paper, we survey these methods; we explain in detail how they work and what their strengths and weaknesses are. We conclude with a research agenda for future work to make the methods more robust and more widely applicable to other applications.

Matthias Müller-Brockhausen, Mike Preuss, Aske Plaat

The idea of transfer in reinforcement learning (TRL) is intriguing: being able to transfer knowledge from one problem to another problem without learning everything from scratch. This promises quicker learning and learning more complex methods. To gain an insight into the field and to detect emerging trends, we performed a database search. We note a surprisingly late adoption of deep learning that starts in 2018. The introduction of deep learning has not yet solved the greatest challenge of TRL: generalization. Transfer between different domains works well when domains have strong similarities (e.g. MountainCar to Cartpole), and most TRL publications focus on different tasks within the same domain that have few differences. Most TRL applications we encountered compare their improvements against self-defined baselines, and the field is still missing unified benchmarks. We consider this to be a disappointing situation. For the future, we note that: (1) A clear measure of task similarity is needed. (2) Generalization needs to improve. Promising approaches merge deep learning with planning via MCTS or introduce memory through LSTMs. (3) The lack of benchmarking tools will be remedied to enable meaningful comparison and measure progress. Already Alchemy and Meta-World are emerging as interesting benchmark suites. We note that another development, the increase in procedural content generation (PCG), can improve both benchmarking and generalization in TRL.

Zhao Yang, Mike Preuss, Aske Plaat

Transfer learning can speed up training in machine learning and is regularly used in classification tasks. It reuses prior knowledge from other tasks to pre-train networks for new tasks. In reinforcement learning, learning actions for a behavior policy that can be applied to new environments is still a challenge, especially for tasks that involve much planning. Sokoban is a challenging puzzle game. It has been used widely as a benchmark in planning-based reinforcement learning. In this paper, we show how prior knowledge improves learning in Sokoban tasks. We find that reusing feature representations learned previously can accelerate learning new, more complex, instances. In effect, we show how curriculum learning, from simple to complex tasks, works in Sokoban. Furthermore, feature representations learned in simpler instances are more general, and thus lead to positive transfers towards more complex tasks, but not vice versa. We have also studied which part of the knowledge is most important for transfer to succeed, and identify which layers should be used for pre-training.

Hui Wang, Mike Preuss, Aske Plaat

AlphaZero has achieved impressive performance in deep reinforcement learning by utilizing an architecture that combines search and training of a neural network in self-play. Many researchers are looking for ways to reproduce and improve results for other games/tasks. However, the architecture is designed to learn from scratch, tabula rasa, accepting a cold-start problem in self-play. Recently, a warm-start enhancement method for Monte Carlo Tree Search was proposed to improve the self-play starting phase. It employs a fixed parameter $I^\prime$ to control the warm-start length. Improved performance was reported in small board games. In this paper we present results with an adaptive switch method. Experiments show that our approach works better than the fixed $I^\prime$, especially for "deep," tactical, games (Othello and Connect Four). We conjecture that the adaptive value for $I^\prime$ is also influenced by the size of the game, and that on average $I^\prime$ will increase with game size. We conclude that AlphaZero-like deep reinforcement learning benefits from adaptive rollout based warm-start, as Rapid Action Value Estimate did for rollout-based reinforcement learning 15 years ago.

Alexander Hagg, Mike Preuss, Alexander Asteroth et al.

More and more, optimization methods are used to find diverse solution sets. We compare solution diversity in multi-objective optimization, multimodal optimization, and quality diversity in a simple domain. We show that multiobjective optimization does not always produce much diversity, multimodal optimization produces higher fitness solutions, and quality diversity is not sensitive to genetic neutrality and creates the most diverse set of solutions. An autoencoder is used to discover phenotypic features automatically, producing an even more diverse solution set with quality diversity. Finally, we make recommendations about when to use which approach.

Christoph Salge, Emily Short, Mike Preuss et al.

Live Action Role-Playing (LARP) games and similar experiences are becoming a popular game genre. Here, we discuss how artificial intelligence techniques, particularly those commonly used in AI for Games, could be applied to LARP. We discuss the specific properties of LARP that make it a surprisingly suitable application field, and provide a brief overview of some existing approaches. We then outline several directions where utilizing AI seems beneficial, by both making LARPs easier to organize, and by enhancing the player experience with elements not possible without AI.

Aske Plaat, Walter Kosters, Mike Preuss

Deep reinforcement learning has shown remarkable success in the past few years. Highly complex sequential decision making problems have been solved in tasks such as game playing and robotics. Unfortunately, the sample complexity of most deep reinforcement learning methods is high, precluding their use in some important applications. Model-based reinforcement learning creates an explicit model of the environment dynamics to reduce the need for environment samples. Current deep learning methods use high-capacity networks to solve high-dimensional problems. Unfortunately, high-capacity models typically require many samples, negating the potential benefit of lower sample complexity in model-based methods. A challenge for deep model-based methods is therefore to achieve high predictive power while maintaining low sample complexity. In recent years, many model-based methods have been introduced to address this challenge. In this paper, we survey the contemporary model-based landscape. First we discuss definitions and relations to other fields. We propose a taxonomy based on three approaches: using explicit planning on given transitions, using explicit planning on learned transitions, and end-to-end learning of both planning and transitions. We use these approaches to organize a comprehensive overview of important recent developments such as latent models. We describe methods and benchmarks, and we suggest directions for future work for each of the approaches. Among promising research directions are curriculum learning, uncertainty modeling, and use of latent models for transfer learning.

Hui Wang, Mike Preuss, Michael Emmerich et al.

Morpion Solitaire is a popular single player game, performed with paper and pencil. Due to its large state space (on the order of the game of Go) traditional search algorithms, such as MCTS, have not been able to find good solutions. A later algorithm, Nested Rollout Policy Adaptation, was able to find a new record of 82 steps, albeit with large computational resources. After achieving this record, to the best of our knowledge, there has been no further progress reported, for about a decade. In this paper we take the recent impressive performance of deep self-learning reinforcement learning approaches from AlphaGo/AlphaZero as inspiration to design a searcher for Morpion Solitaire. A challenge of Morpion Solitaire is that the state space is sparse, there are few win/loss signals. Instead, we use an approach known as ranked reward to create a reinforcement learning self-play framework for Morpion Solitaire. This enables us to find medium-quality solutions with reasonable computational effort. Our record is a 67 steps solution, which is very close to the human best (68) without any other adaptation to the problem than using ranked reward. We list many further avenues for potential improvement.

Jialin Liu, Antoine Moreau, Mike Preuss et al.

Choosing automatically the right algorithm using problem descriptors is a classical component of combinatorial optimization. It is also a good tool for making evolutionary algorithms fast, robust and versatile. We present Shiwa, an algorithm good at both discrete and continuous, noisy and noise-free, sequential and parallel, black-box optimization. Our algorithm is experimentally compared to competitors on YABBOB, a BBOB comparable testbed, and on some variants of it, and then validated on several real world testbeds.

Hui Wang, Mike Preuss, Aske Plaat

Recently, AlphaZero has achieved landmark results in deep reinforcement learning, by providing a single self-play architecture that learned three different games at super human level. AlphaZero is a large and complicated system with many parameters, and success requires much compute power and fine-tuning. Reproducing results in other games is a challenge, and many researchers are looking for ways to improve results while reducing computational demands. AlphaZero's design is purely based on self-play and makes no use of labeled expert data ordomain specific enhancements; it is designed to learn from scratch. We propose a novel approach to deal with this cold-start problem by employing simple search enhancements at the beginning phase of self-play training, namely Rollout, Rapid Action Value Estimate (RAVE) and dynamically weighted combinations of these with the neural network, and Rolling Horizon Evolutionary Algorithms (RHEA). Our experiments indicate that most of these enhancements improve the performance of their baseline player in three different (small) board games, with especially RAVE based variants playing strongly.

Marco Pleines, Jenia Jitsev, Mike Preuss et al.

The Obstacle Tower Challenge is the task to master a procedurally generated chain of levels that subsequently get harder to complete. Whereas the most top performing entries of last year's competition used human demonstrations or reward shaping to learn how to cope with the challenge, we present an approach that performed competitively (placed 7th) but starts completely from scratch by means of Deep Reinforcement Learning with a relatively simple feed-forward deep network structure. We especially look at the generalization performance of the taken approach concerning different seeds and various visual themes that have become available after the competition, and investigate where the agent fails and why. Note that our approach does not possess a short-term memory like employing recurrent hidden states. With this work, we hope to contribute to a better understanding of what is possible with a relatively simple, flexible solution that can be applied to learning in environments featuring complex 3D visual input where the abstract task structure itself is still fairly simple.

Matthias Muller-Brockhausen, Mike Preuss, Aske Plaat

Decades of research have been invested in making computer programs for playing games such as Chess and Go. This paper focuses on a new game, Tetris Link, a board game that is still lacking any scientific analysis. Tetris Link has a large branching factor, hampering a traditional heuristic planning approach. We explore heuristic planning and two other approaches: Reinforcement Learning, Monte Carlo tree search. We document our approach and report on their relative performance in a tournament. Curiously, the heuristic approach is stronger than the planning/learning approaches. However, experienced human players easily win the majority of the matches against the heuristic planning AIs. We, therefore, surmise that Tetris Link is more difficult than expected. We offer our findings to the community as a challenge to improve upon.

Lena Clever, Dennis Assenmacher, Kilian Müller et al.

Nowadays fake news are heavily discussed in public and political debates. Even though the phenomenon of intended false information is rather old, misinformation reaches a new level with the rise of the internet and participatory platforms. Due to Facebook and Co., purposeful false information - often called fake news - can be easily spread by everyone. Because of a high data volatility and variety in content types (text, images,...) debunking of fake news is a complex challenge. This is especially true for automated approaches, which are prone to fail validating the veracity of the information. This work focuses on an a gamified approach to strengthen the resilience of consumers towards fake news. The game FakeYou motivates its players to critically analyze headlines regarding their trustworthiness. Further, the game follows a "learning by doing strategy": by generating own fake headlines, users should experience the concepts of convincing fake headline formulations. We introduce the game itself, as well as the underlying technical infrastructure. A first evaluation study shows, that users tend to use specific stylistic devices to generate fake news. Further, the results indicate, that creating good fakes and identifying correct headlines are challenging and hard to learn.

Hui Wang, Michael Emmerich, Mike Preuss et al.

The landmark achievements of AlphaGo Zero have created great research interest into self-play in reinforcement learning. In self-play, Monte Carlo Tree Search is used to train a deep neural network, that is then used in tree searches. Training itself is governed by many hyperparameters.There has been surprisingly little research on design choices for hyper-parameter values and loss-functions, presumably because of the prohibitive computational cost to explore the parameter space. In this paper, we investigate 12 hyper-parameters in an AlphaZero-like self-play algorithm and evaluate how these parameters contribute to training. We use small games, to achieve meaningful exploration with moderate computational effort. The experimental results show that training is highly sensitive to hyper-parameter choices. Through multi-objective analysis we identify 4 important hyper-parameters to further assess. To start, we find surprising results where too much training can sometimes lead to lower performance. Our main result is that the number of self-play iterations subsumes MCTS-search simulations, game-episodes, and training epochs. The intuition is that these three increase together as self-play iterations increase, and that increasing them individually is sub-optimal. A consequence of our experiments is a direct recommendation for setting hyper-parameter values in self-play: the overarching outer-loop of self-play iterations should be maximized, in favor of the three inner-loop hyper-parameters, which should be set at lower values. A secondary result of our experiments concerns the choice of optimization goals, for which we also provide recommendations.

Sebastian Risi, Mike Preuss

This paper reviews the field of Game AI, which not only deals with creating agents that can play a certain game, but also with areas as diverse as creating game content automatically, game analytics, or player modelling. While Game AI was for a long time not very well recognized by the larger scientific community, it has established itself as a research area for developing and testing the most advanced forms of AI algorithms and articles covering advances in mastering video games such as StarCraft 2 and Quake III appear in the most prestigious journals. Because of the growth of the field, a single review cannot cover it completely. Therefore, we put a focus on important recent developments, including that advances in Game AI are starting to be extended to areas outside of games, such as robotics or the synthesis of chemicals. In this article, we review the algorithms and methods that have paved the way for these breakthroughs, report on the other important areas of Game AI research, and also point out exciting directions for the future of Game AI.

Marwin H. S. Segler, Mike Preuss, Mark P. Waller

From medicines to materials, small organic molecules are indispensable for human well-being. To plan their syntheses, chemists employ a problem solving technique called retrosynthesis. In retrosynthesis, target molecules are recursively transformed into increasingly simpler precursor compounds until a set of readily available starting materials is obtained. Computer-aided retrosynthesis would be a highly valuable tool, however, past approaches were slow and provided results of unsatisfactory quality. Here, we employ Monte Carlo Tree Search (MCTS) to efficiently discover retrosynthetic routes. MCTS was combined with an expansion policy network that guides the search, and an "in-scope" filter network to pre-select the most promising retrosynthetic steps. These deep neural networks were trained on 12 million reactions, which represents essentially all reactions ever published in organic chemistry. Our system solves almost twice as many molecules and is 30 times faster in comparison to the traditional search method based on extracted rules and hand-coded heuristics. Finally after a 60 year history of computer-aided synthesis planning, chemists can no longer distinguish between routes generated by a computer system and real routes taken from the scientific literature. We anticipate that our method will accelerate drug and materials discovery by assisting chemists to plan better syntheses faster, and by enabling fully automated robot synthesis.

Simon Wessing, Mike Preuss

Efficient global optimization is a popular algorithm for the optimization of expensive multimodal black-box functions. One important reason for its popularity is its theoretical foundation of global convergence. However, as the budgets in expensive optimization are very small, the asymptotic properties only play a minor role and the algorithm sometimes comes off badly in experimental comparisons. Many alternative variants have therefore been proposed over the years. In this work, we show experimentally that the algorithm instead has its strength in a setting where multiple optima are to be identified.